This research project focuses on improving Near-Earth Object (NEO) detection using advanced machine learning techniques, particularly Vision Transformers (ViTs). The study addresses challenges such as noise, limited data, and class imbalance. The ViT model, initially designed for natural language tasks, has been adapted for image processing to effectively capture complex patterns and relationships in astronomical data. The methodology involved preparing a curated dataset of NEO images, resizing them to 128x128 pixels, and organizing them into triplet sequences. ViTs processed these sequences, leveraging self-attention and feed-forward neural networks (FFNNs) to distinguish NEOs from other objects as well as track the NEO’s trajectory. Multiple learning rates and batch sizes were tested, revealing the optimal combinations for stability and accuracy. The results revealed distinct behaviors associated with varying learning rates. Notably, the learning rate of 0.001 consistently demonstrated stable convergence in training and high accuracy in testing across different batch sizes. In contrast, a learning rate of 0.01 exhibited significant fluctuations in the loss function, indicating challenges in training stability. Conversely, a learning rate of 0.0001 showcased relatively low and consistent loss values during training. These insights highlight the potential of the ViT model for enhancing NEO detection by effectively capturing temporal and spatial patterns. Furthermore, the study emphasizes the significance of larger and more diverse datasets, addressing class imbalances, and enhancing model transparency for guiding future research. In summary, ViTs hold the potential to enhance NEO detection by shedding light on the dynamics of celestial objects and contributing to planetary defense initiatives. The knowledge gained from parameter exploration serves as valuable guidance for optimizing ViT models for NEO detection. Moreover, continuous advancements in NEO detection techniques pave the way for the discovery of previously unknown celestial entities. / Detta forskningsprojekt fokuserar på att förbättra detektering av Near-Earth Object (NEO) med hjälp av avancerad maskininlärningsteknik, särskilt Vision Transformers (ViTs). Studien tar upp utmaningar som buller, begränsad data och klassobalans. ViT-modellen, från början designad för naturliga språkuppgifter, har anpassats för bildbehandling för att effektivt fånga komplexa mönster och samband i astronomiska data. Metodiken innebar att förbereda en kurerad datauppsättning av NEO-bilder, ändra storlek på dem till 128x128 pixlar och organisera dem i triplettsekvenser. ViTs bearbetade dessa sekvenser, utnyttjade självuppmärksamhet och feedforward neurala nätverk (FFNNs) för att skilja NEOs från andra objekt samt spåra NEO’s bana. Flera inlärningshastigheter och batchstorlekar testades, vilket avslöjade de optimala kombinationerna för stabilitet och noggrannhet. Resultaten avslöjade distinkta beteenden associerade med varierande inlärningshastigheter. Noterbart visade inlärningshastigheten på 0,001 konsekvent stabil konvergens i träning och hög noggrannhet i testning över olika batchstorlekar. Däremot uppvisade en inlärningshastighet på 0,01 signifikanta fluktuationer i förlustfunktionen, vilket indikerar utmaningar i träningsstabilitet. Omvänt visade en inlärningshastighet på 0,0001 relativt låga och konsekventa förlustvärden under träning. Dessa insikter belyser potentialen hos ViT-modellen för att förbättra NEO-detektering genom att effektivt fånga tids- och rumsmönster. Dessutom betonar studien betydelsen av större och mer varierande datauppsättningar, tar itu med klassobalanser och förbättrar modelltransparensen för att vägleda framtida forskning.svis har ViTs potentialen att förbättra NEO-detektering genom att belysa dynamiken hos himmelska objekt och bidra till planetariska försvarsinitiativ. Kunskapen från parameterutforskning fungerar som värdefull vägledning för att optimera ViT-modeller för NEO-detektering. Dessutom banar kontinuerliga framsteg inom NEO-detektionstekniker vägen för upptäckten av tidigare okända himmelska entiteter.

In this thesis, the possibility of enabling space-mining for the upcoming decade is explored. Making use of recently-proven reusable rockets, we envision a fleet of spacecraft capable of reaching Near-Earth asteroids. To analyze this idea, the goal of this problem is to maximize the asteroid mass retrieved within a spacecraft max life span. Explicitly, the maximum lifetime of the spacecraft fleet is set at 30 years. A fuel supply-chain is proposed and designed so that each spacecraft is refueled before departing for each asteroid. To maximize access to the number of asteroids and retrievable mass for each mission, we propose launching each mission from an orbit with low escape velocity. The L2-Halo orbit at the libration point in the Earth-Moon system was selected due to its easy access from Low-Earth Orbit and for a cislunar synergy with NASA Gateway. Using data from NASA SmallBody and CNEOS databases, we investigated NEAs in the period between 2030 and 2060 could be captured in the ecliptic plane and returned to L2-Halo with two approaches, MARM-1 and MARM-2. Together, these databases provide all information for every asteroid's close approach known today. Returning the asteroid as a whole is explored in the MARM-1 method, while MARM-2 evaluates the possibility of reaching larger asteroids and returning a fragment of their masses, such that it optimizes the available cargo weight per time of flight of each mission. The following results are compared with previous work from the community. The results show a 96% reduction in the cost per kg, with an enormous increase in retrieved mass. With these results, this thesis shows that not solely energy or dynamic optimization will be responsible for proving space mining feasibility, but rather a combination of those and business best practices. Proving feasibility for space mining is a complex and immense problem. Although this thesis opens new possibilities for future work on the field and sparkes the interest of private endeavors, the final solution for this problem still requires additional exploration. / M.S. / In this thesis, the possibility of enabling space-mining for the upcoming decade is explored. Making use of recently-proven reusable rockets, we envision a fleet of spacecraft capable of reaching Near-Earth asteroids, NEAs. To analyze this idea, the goal of this problem is to maximize the asteroid mass retrieved within a spacecraft max life span. Explicitly, the maximum lifetime of the spacecraft fleet is set at 30 years. A fuel supply-chain is proposed and designed so that each spacecraft is refueled before departing for each asteroid. To maximize access to the number of asteroids and retrievable mass for each mission, we propose launching each mission from an orbit with low escape velocity. A location after the Moon, at the L2-Halo orbit, was selected due to its easy access from Low-Earth Orbit and for a synergy with the proposed new space station at the Moon orbit. Using data from NASA databases, we investigated the asteroids in the period between 2030 and 2060 that could be captured and returned with two approaches, MARM-1 and MARM-2. Together, these databases provide all information for every asteroid's close approach known today. Returning the asteroid as a whole is explored in the MARM-1 method, while MARM-2 evaluates the possibility of reaching larger asteroids and returning a fragment of their masses, such that it optimizes the available cargo weight per time of flight of each mission. The following results are compared with previous work from the community. The results show a 96% reduction in the cost per kg, with an enormous increase in retrieved mass. With these results, this thesis shows that not solely energy or dynamic optimization will be responsible for proving space mining feasibility, but rather a combination of those and business best practices. Proving feasibility for space mining is a complex and immense problem. Although this thesis opens new possibilities for future work on the field and sparkes the interest of private endeavors, the final solution for this problem still requires additional exploration.